Voice Analyze

voice-analyze

Reverse-engineer voice profiles from sample content by analyzing writing patterns. Detects and flags AI writing tells so the resulting profile produces human-sounding output.

Triggers

Alternate expressions and non-obvious activations (primary phrases are matched automatically from the skill description):

• "voice fingerprint" -> voice profile extraction
• "what's my writing style" -> voice analysis

Behavior

When triggered, this skill:

1. Analyzes text samples for:

   • Sentence structure and length patterns
   • Vocabulary sophistication and domain
   • Tone markers (formality, confidence, warmth)
   • Structural patterns (lists, examples, questions)
   • Perspective and voice choices

2. Extracts measurable features:

   • Average sentence length
   • Vocabulary complexity (syllables, word length)
   • Contraction usage
   • Personal pronoun frequency
   • Question density
   • List/bullet usage

3. Scans for AI writing tells (see ../voice-apply/references/ai-tells.md):

   • Flags any banned vocabulary found in the sample (verbs, adjectives, nouns, adverbs from the tells list)
   • Detects banned sentence structures: "It's not X, it's Y" negation patterns, self-posed questions answered immediately, anaphora abuse, present participle trailing clauses
   • Measures em dash frequency (AI models use ~10x more than human writers)
   • Checks for uniform sentence length clustering (the "burstiness problem": AI clusters at 15-20 words, humans swing from 3-word fragments to 40-word sprawls)
   • Detects banned transitions and filler phrases (throat-clearing openers, pedagogical openers, fake-suspense transitions, hype phrases)
   • Flags formatting tells: bold-first bullet points, excessive headers, erratic bolding, emoji decoration
   • Checks for Rule of Three overuse (back-to-back tricolons are an AI fingerprint)
   • Detects uniform paragraph length (the most visually obvious tell)
   • If the sample contains AI tells, the profile's vocabulary.avoid list MUST include them and the ai_tells_report section must document what was found

4. Maps features to voice dimensions:

   • Statistical analysis -> tone scale values (0-1)
   • Pattern detection -> structure preferences
   • Vocabulary extraction -> prefer/avoid lists

5. Generates voice profile matching the analyzed style, with AI tells filtered out

Usage Examples

Analyze Existing Documentation

User: "Analyze this writing style" + [paste technical docs]

Analysis:
- Formality: 0.7 (no contractions, structured sentences)
- Confidence: 0.85 (direct statements, few hedges)
- Warmth: 0.25 (impersonal, third-person)
- Complexity: 0.8 (technical vocabulary, long sentences)
- AI Tells Detected: 3 (2 banned verbs, 1 negation pattern)

Output: analyzed-technical-docs.yaml

Match Brand Voice

User: "Extract voice from our marketing copy" + [paste samples]

Analysis:
- Formality: 0.3 (conversational, contractions)
- Confidence: 0.7 (benefit claims, but some hedging)
- Warmth: 0.85 (second person, friendly tone)
- Energy: 0.8 (exclamation points, action verbs)
- AI Tells Detected: 0

Output: brand-marketing-voice.yaml

Capture Personal Style

User: "Create profile from my blog posts" + [paste samples]

Analysis:
- Identifies personal writing quirks
- Extracts signature phrases
- Maps to voice dimensions
- Flags and excludes any AI tells found in samples

Output: personal-blog-voice.yaml

Analysis Methodology

Feature Extraction

AI Tells Detection Pass

After standard feature extraction, run the AI tells scan. For each category in ../voice-apply/references/ai-tells.md:

1. Vocabulary scan: Count occurrences of every word/phrase in the banned lists (verbs, adjectives, nouns, adverbs, connectors)
2. Structure scan: Detect banned sentence patterns (negation flips, self-posed questions, anaphora, trailing participles, false ranges, hedge-stacking)
3. Punctuation scan: Count em dashes, semicolons, Oxford commas, Unicode ellipsis characters
4. Rhythm scan: Calculate sentence length standard deviation. If std dev < 4 words, flag as AI-like clustering
5. Formatting scan: Check for bold-first bullets, excessive headers, erratic bolding
6. Transition scan: Search for throat-clearing openers, pedagogical openers, fake-suspense phrases, hype phrases, banned conclusion phrases

Report findings in the output profile under ai_tells_report. Any banned vocabulary found in the sample goes into vocabulary.avoid automatically.

Dimension Calibration

Formality (0-1):

• 0.0-0.3: Contractions frequent, casual language, fragments okay
• 0.4-0.6: Mixed style, professional but accessible
• 0.7-1.0: No contractions, complete sentences, formal structure

Confidence (0-1):

• 0.0-0.3: Many hedges ("might", "perhaps"), questions, qualifiers
• 0.4-0.6: Balanced certainty, occasional hedges
• 0.7-1.0: Direct statements, conclusions first, few qualifiers

Warmth (0-1):

• 0.0-0.3: Third person, passive voice, clinical tone
• 0.4-0.6: Professional but personable
• 0.7-1.0: Second person, inclusive language, empathetic

Energy (0-1):

• 0.0-0.3: Calm, measured, understated
• 0.4-0.6: Balanced engagement
• 0.7-1.0: Exclamation marks, action verbs, dynamic phrasing

Complexity (0-1):

• 0.0-0.3: Short sentences, simple vocabulary, accessible
• 0.4-0.6: Moderate complexity, clear but nuanced
• 0.7-1.0: Long sentences, technical vocabulary, layered ideas

Vocabulary Extraction

Signature phrases - Identified by:

• Repeated patterns across samples
• Distinctive constructions
• Opening/closing patterns

Domain vocabulary - Extracted by:

• Technical term frequency
• Specialized jargon
• Industry-specific language

Avoid patterns - Built from two sources:

1. Conspicuous absence of common phrases in the samples
2. Any AI writing tells found during the tells detection pass (always include these)

Output Format

name: analyzed-sample-voice
version: 1.0.0
description: Voice profile extracted from sample content
analysis_source:
  sample_size: 1500  # words analyzed
  sample_count: 3    # number of samples
  confidence: 0.85   # analysis confidence score
tone:
  formality: 0.65
  confidence: 0.8
  warmth: 0.4
  energy: 0.5
  complexity: 0.7
vocabulary:
  prefer:
    - "extracted signature phrase 1"
    - "detected domain terminology"
  avoid:
    - "leverage"       # AI tell: banned verb
    - "delve"          # AI tell: banned verb
    - "comprehensive"  # AI tell: banned adjective
    - "landscape"      # AI tell: banned noun
    - "furthermore"    # AI tell: banned adverb
    - "patterns not found in samples"
  signature_phrases:
    - "The key point is..."
    - "This demonstrates..."
structure:
  sentence_length: medium    # avg 15-20 words
  sentence_length_variance: high  # std dev > 6 words (human-like)
  paragraph_length: medium   # avg 4-6 sentences
  paragraph_length_variance: high # varied paragraph sizes (human-like)
  sentence_variety: high     # varied structure detected
  use_lists: when-appropriate
  use_examples: frequently
  use_questions: rarely
  em_dash_usage: avoid       # always set to avoid unless user explicitly uses them naturally
perspective:
  person: third
  voice: active
  tense: present
extracted_patterns:
  opening_style: "context-first"
  closing_style: "conclusion-summary"
  transition_style: "logical-flow"
ai_tells_report:
  tells_found: 3
  em_dash_frequency: 2.1     # per 1000 words (human avg ~0.5, AI avg ~5.0)
  sentence_length_std_dev: 7.2  # words (human = 6-12, AI = 2-4)
  paragraph_length_std_dev: 2.8 # sentences (human = 2-5, AI = 0.5-1.5)
  banned_vocabulary_found:
    - "leverage (1x)"
    - "comprehensive (2x)"
  banned_structures_found:
    - "negation pattern (1x)"
  banned_transitions_found: []
  formatting_tells_found: []
  overall_human_score: 0.82  # 0 = clearly AI, 1 = clearly human

Integration

• Output: Creates profiles usable by voice-apply
• Chain: voice-analyze -> voice-create (to refine) -> voice-apply
• Chain: voice-analyze + voice-analyze -> voice-blend (combine styles)

Accuracy Considerations

• Minimum sample: 500+ words for reliable analysis
• Multiple samples: 3+ samples improve accuracy
• Consistent genre: Mixing genres reduces accuracy
• Confidence score: Output includes analysis confidence (0-1)
• AI tells in samples: If the sample itself contains AI tells (e.g., it was partially AI-generated), the analyzer will flag them and exclude them from the profile's preferred patterns. The overall_human_score helps gauge how much of the sample is genuinely the user's voice vs AI artifacts.

References

• AI Writing Tells: ../voice-apply/references/ai-tells.md

Source: angelarose210/ghostwriter — distributed by TomeVault.